[Autoscaler] Proper resource demand plumbing (#10329)

python/ray/autoscaler/autoscaler.py

@@ -117,7 +117,7 @@ class StandardAutoscaler:
         # Aggregate resources the user is requesting of the cluster.
         self.resource_requests = defaultdict(int)
         # List of resource bundles the user is requesting of the cluster.
-        self.resource_demand_vector = None
+        self.resource_demand_vector = []
 
         logger.info("StandardAutoscaler: {}".format(self.config))
 
@@ -197,14 +197,18 @@ class StandardAutoscaler:
             self.log_info_string(nodes, target_workers)
 
         # First let the resource demand scheduler launch nodes, if enabled.
-        if self.resource_demand_scheduler and self.resource_demand_vector:
-            to_launch = (self.resource_demand_scheduler.get_nodes_to_launch(
-                self.provider.non_terminated_nodes(tag_filters={}),
-                self.pending_launches.breakdown(),
-                self.resource_demand_vector))
-            # TODO(ekl) also enforce max launch concurrency here?
-            for node_type, count in to_launch:
-                self.launch_new_node(count, node_type=node_type)
+        if self.resource_demand_scheduler:
+            resource_demand_vector = self.resource_demand_vector + \
+                self.load_metrics.get_resource_demand_vector()
+            if resource_demand_vector:
+                to_launch = (
+                    self.resource_demand_scheduler.get_nodes_to_launch(
+                        self.provider.non_terminated_nodes(tag_filters={}),
+                        self.pending_launches.breakdown(),
+                        resource_demand_vector))
+                # TODO(ekl) also enforce max launch concurrency here?
+                for node_type, count in to_launch:
+                    self.launch_new_node(count, node_type=node_type)
 
         # Launch additional nodes of the default type, if still needed.
         num_workers = len(nodes) + num_pending

python/ray/autoscaler/load_metrics.py

@@ -23,8 +23,16 @@ class LoadMetrics:
         self.dynamic_resources_by_ip = {}
         self.resource_load_by_ip = {}
         self.local_ip = services.get_node_ip_address()
+        self.waiting_bundles = []
+        self.infeasible_bundles = []
 
-    def update(self, ip, static_resources, dynamic_resources, resource_load):
+    def update(self,
+               ip,
+               static_resources,
+               dynamic_resources,
+               resource_load,
+               waiting_bundles=[],
+               infeasible_bundles=[]):
         self.resource_load_by_ip[ip] = resource_load
         self.static_resources_by_ip[ip] = static_resources
 
@@ -43,6 +51,8 @@ class LoadMetrics:
                 static_resources != dynamic_resources:
             self.last_used_time_by_ip[ip] = now
         self.last_heartbeat_time_by_ip[ip] = now
+        self.waiting_bundles = waiting_bundles
+        self.infeasible_bundles = infeasible_bundles
 
     def mark_active(self, ip):
         assert ip is not None, "IP should be known at this time"
@@ -127,6 +137,9 @@ class LoadMetrics:
 
         return nodes_used, resources_used, resources_total
 
+    def get_resource_demand_vector(self):
+        return self.waiting_bundles + self.infeasible_bundles
+
     def info_string(self):
         return " - " + "\n - ".join(
             ["{}: {}".format(k, v) for k, v in sorted(self._info().items())])